基于MAC-LSTM的问题分类研究

针对问句文本通常较短、语义信息与词语共现信息不足等问题，提出一种多层级注意力卷积长短时记忆模型（multi-level attention convolution LSTM neural network，MAC-LSTM）的问题分类方法。相比基于词嵌入的深度学习模型，该方法使用疑问词注意力机制对问句中的疑问词特征重点关注。同时，使用注意力机制结合卷积神经网络与长短时记忆模型各自文本建模的优势，既能够并行方式提取词汇级特征，又能够学习更高级别的长距离依赖特征。实验表明，该方法较传统的机器学习方法和普通的卷积神经网络、长短时记忆模型有明显的效果提升。

Question classification based on MAC-LSTM

Question text is usually short and the information of semantic information and word co-occurrence are not enough. To address the above problems, this paper proposed a multi-level attention convolution LSTM neural network(MAC-LSTM) for question classification. This approach used the interrogative word attention mechanism to focus on the interrogative features in the heterogeneous question contexts. At the same time, it used the attention mechanism combined with the advantages of convolutional neural network and long-short memory model recurrent neural network(LSTM). MAC-LSTM is able to capture both local features of phrases as well as global and temporal sentence semantics. Experiments show that, this approach achieves better performance than traditional machine learning method, ordinary convolutional neural network, and traditional LSTM on question classification tasks without the need of prior knowledge.